Machine learning prediction of physical properties of lignin derived porous carbon via catalytic pyrolysis

Abstract

Lignin-derived porous carbon produced through catalytic pyrolysis is crucial for energy storage, adsorption, and catalysis. However, predicting specific surface area (SSA), total pore volume (TPV), and microporosity (MP) remains challenging due to the variability in lignin properties, chemical activators, and pyrolysis conditions, compounded by limited data availability. In this study, we applied a hybrid machine learning framework incorporating a pre-trained interpolation model and a final regressor to impute missing features, improving prediction accuracy and generalizability. This approach yielded high predictive accuracy with R² values of 0.82 (SSA), 0.86 (TPV), and 0.81 (MP) on a dataset of 112 samples, encompassing variations across six chemical activators (KOH, ZnCl₂, H₃PO₄, K₂CO₃, NaOH, and Na₂CO₃). Feature importance analysis highlighted the significant influence of KOH on SSA and TPV, and H₃PO₄ on MP. This research provides a framework to precisely tailor the pore structure of lignin-derived porous carbon via catalytic pyrolysis, enabling advancements in applications across diverse fields.